Open-ended Evaluation on RadBench open-ended reduced split n=69
2RadGraph F1HR1.5
Evaluation Results
| Method | Links | |||||||
|---|---|---|---|---|---|---|---|---|
| HR1.52026.07 | 2 | 13.3 | 73.9 | 85.8 | 4.9 | 47 | 37.1 | |
| HR1.5+2026.07 | 0.8 | 27.4 | 58 | 79.3 | 1.7 | 20.1 | 31 | |
| Gemini-3-Flash-Preview2026.07 | 0.7 | 1.8 | 71 | 92.1 | 8.5 | 60.8 | 35.1 | |
| GPT-5.42026.07 | 0.6 | 11.2 | 78.3 | 88.5 | 5.3 | 49.7 | 33.3 | |
| Claude Opus 4.72026.07 | 0.6 | 46 | 31.9 | 65.4 | 0.2 | 6.5 | 24.9 | |
| HR12026.07 | 0.3 | 28.7 | 59.4 | 78.8 | 0 | 17.3 | 28.4 | |
| MedGemma-27B2026.07 | 0 | 14.9 | 59.4 | 87.2 | 3.6 | 50.6 | 12.8 | |
| MedGemma-4B2026.07 | 0 | 12.2 | 58 | 88.6 | 0 | 48.9 | 16.5 |